Serverless AI inference

Send supported chat, image, or video requests through an API with scoped credentials and wallet authorization. Discover available models before making a request.

Choose a supported request, not a virtual machine

Serverless AI provides model requests without an SSH instance to administer. Start with the current model catalog, select its public model ID and modality, then integrate the matching endpoint. For custom checkpoints, operating-system access or a serving protocol you control, compare dedicated GPU operations instead.

Discover, validate, then authorize generation

  1. Prepare an active, email-verified account and a backend API key. Model discovery requires pricing:read; generation requires inference:write. Keep the bearer key out of browser code.
  2. Call GET https://serverless.hostnotgpu.ae/v1/models. Inspect the returned data rows: public ID, modality, context length and pricing. Discovery does not reserve capacity.
  3. Check the exact request schema in the API reference. Choose bounded input and output, review wallet availability and send one logical operation with an Idempotency-Key.
  4. Validate HTTP status and the completed response before displaying output. Record request ID, time and billing evidence without logging private prompts or keys.

Current capabilities and integration limits

Chat uses POST /v1/inference/chat with string messages and returns completed JSON: object: inference.result, with text in output.text. Streaming is not supported: stream: true is rejected. Unknown fields, including tool and structured-output options, are rejected. An OpenAI-shaped messages array is not drop-in SDK compatibility.

Image and video have separate schemas and depend on available backend capabilities and the selected model. Chat and image use normalized results; video follows a separate response path. The model catalog does not expose streaming or tool capability flags. Do not infer endpoint support from an upstream model card or a fixed model count.

Understand wallet and failure states

Generation reserves an estimate before dispatch and settles delivered work. Available funds exclude reservations. A client timeout does not prove that work stopped or that nothing was charged. The operation key supports reconciliation but is not a general cached-result replay guarantee. Prevent concurrent duplicate submissions and investigate uncertain delivery before retrying.

Handle 429 with bounded admission and Retry-After when present; correct authentication, schema, model and balance errors before resubmission. GET /v1/usage is GPU-instance usage, not a combined inference export. Use generation responses and the appropriate workspace records for inference accounting.

Build a small observable integration

Follow the serverless quickstart for a bounded first request, failure handling for uncertain outcomes and the architecture comparison for workload fit. Establish output quality, latency, concurrency and spending limits before scaling; no fixed throughput or availability is promised by catalog visibility.